Framework and Layer-Wise Word-Line Activation Method Design for CIM

Convolution Neural Network (CNN) has excellent performance in various fields such as machine learning, computer vision, and image recognition. With the development of CNNs, huge amount of data in computing and transmission has imposed significant pressure in circuit and architecture design, and RRAM based computing-in-memory (CIM) is one of the promising solutions to alleviate this problem. However, because of the current deviation phenomenon and the resistance on/off ratio (R ratio) issue in RRAM, there is a tradeoff problem between computing accuracy and efficiency for CIM. In this paper, we propose a layer-wise word-line activation strategy to configure the appropriate number of word-line activations for each layer. Based on the observed risk factors, we design a risk index to activated word-lines mapping methodology. Meanwhile, based on the proposed quantization and error calculation methods, we design a CIM simulation framework to simulate the accuracy of CNNs in the inference stage. Experiment results show that for small R ratio, our proposed layer-wise word-line activation methodology reduces up to 63% and 56% of the computation cycles for Cifar-10 and VGG-8, respectively; while the accuracy is decreased by only 0.65% and 0.11%, respectively. On the exploration of increasing the size of R ratio, experiment results show that large R ratio allows our methodology to have a more flexible planning space and better computational efficiency.